Agents API

What it is

Where a basic chat completion endpoint returns text for one prompt, an agents interface manages a loop: the model decides which tool to call, receives the result and continues until the task is done. Typical features include hosted tools such as web search and file search, function calling into your own systems, memory or state across turns, handoffs between specialised agents, and tracing for debugging. Guardrails and permissions are usually layered on so the agent cannot take actions outside a defined scope.

Why it matters

Agentic systems are how assistants move from answering questions to doing work, which changes both product design and discovery. If agents will retrieve and act on your content, then machine readable pages, clean APIs and structured product data become part of how you get chosen. For marketing teams, the same APIs power internal agents that handle research, reporting and content operations at a scale manual work cannot match.

How it works

Developers define an agent with instructions, a set of tools and limits, then run it against tasks while logging each step through tracing. Practitioners typically start with a narrow, high volume task, add evaluation cases, and only widen scope once accuracy and cost are understood. Retrieval over your own documents is commonly bolted on so answers cite internal sources rather than relying on model memory.

When it applies

It applies when a task needs several steps, live data or actions in other systems, and when you need auditable logs of what the assistant did.

Examples

• A marketing team builds an agent that pulls Search Console and analytics data each Monday, then drafts a commentary summary for review.
• A support product runs an agent that searches help documentation, checks the customer's plan through an internal function, then drafts a reply.
• An ecommerce brand exposes a stock and delivery API so third party agents can answer availability questions accurately.

How it is measured

• Task completion rate without human correction
• Tool call accuracy and error rate per step
• Cost and latency per completed task
• Human review or escalation rate over time